## A Filter Active-Set Trust-Region Method (2007)

### Cached

### Download Links

Citations: | 3 - 0 self |

### BibTeX

@MISC{Friedl07afilter,

author = {Michael P. Friedl and Nick I. M. Gould and Sven Leyffer and Todd S. Munson},

title = {A Filter Active-Set Trust-Region Method},

year = {2007}

}

### OpenURL

### Abstract

2.1 Sequential Linear-Quadratic Programming Methods.............. 3 2.2 Difficulties with the LP/TR Step Computation................ 4

### Citations

915 |
Solution of Ill-Posed Problems
- Tikhonov, Arsenin
- 1977
(Show Context)
Citation Context ...e return to this issue below, but assume for now that one can find a solution dLP to RLP(x, µ). The problem RLP(x, µ) has an interesting history. In particular, RLP(x, µ) is a Tikhonov regularization =-=[27]-=- of LP(x), for which the solution of the former is the least ℓ2-norm solution to the latter for all µ sufficiently large [22]. See [29] for further references. Vitally, since the Hessian of the object... |

178 | Nonlinear programming without a penalty function
- Fletcher, Leyffer
- 2002
(Show Context)
Citation Context ...∆f = f(x) − f(x + d) be the actual reduction in f(x) resulting from a given step d. We measure constraint infeasibility by h(c(x)) = �c(x) + �∞, where ci(x) + = max(0, −ci(x)). We now define a filter =-=[10]-=- that combines the two competing aims in NLP, namely, minimization of f(x) and h(c(x)). We will consider pairs (h, f) obtained by evaluating f(x) and h(c(x)). We say that a pair (hk, fk) dominates ano... |

131 |
Application of second-order cone programming, Linear Algebra and its
- Lobo, Vandenberghe, et al.
- 1998
(Show Context)
Citation Context ... appropriate, but the downside is that TR(x, ρ) is no longer an LP and may be more difficult to solve—there has been some work for the ℓ2-norm case using semi-definite and cone programming techniques =-=[1, 12, 17, 18, 20, 26]-=-, but to our knowledge the size of problems that are amenable to such techniques fall far short of those currently possible for LP subproblems. A second difficulty, common to many trust-region methods... |

126 |
A version of the bundle idea for minimizing a nonsmooth function: conceptual idea, convergence analysis, numerical results
- Schramm, Zowe
- 1992
(Show Context)
Citation Context ...atter for all µ sufficiently large [22]. See [29] for further references. Vitally, since the Hessian of the objective is diagonal, the dual of RLP(x, µ) is a bound-constrained quadratic program (BQP) =-=[16, 21]-=- in the dual variables y, (BQP(x, µ)) � minimize y subject to y ≥ 0, 1 2yT AT (x)A(x)y + � c(x) − µAT (x)g(x) �T µ y + 2 2 gT (x)g(x) from which the solution d = −µg(x)+A(x)y of RLP(x, µ) may be recov... |

100 |
LANCELOT: A Fortran Package for Large-Scale Nonlinear Optimization (Release A
- Toint
- 1992
(Show Context)
Citation Context ... � c(x) − µAT (x)g(x) �T µ y + 2 2 gT (x)g(x) from which the solution d = −µg(x)+A(x)y of RLP(x, µ) may be recovered. The significance is that there are excellent methods [7, 15, 21, 23] and software =-=[8, 19]-=- for solving large-scale BQPs. Having solved RLP(x, µ)—or its dual, BQP(x, µ)—to find an initial step dLP and its associated active set A, our algorithm solves the resulting EQP(x, A) to find a second... |

97 | Interior-point methods for nonconvex nonlinear programming: orderings and higher-order methods
- Shanno, Vanderbei
(Show Context)
Citation Context ...m an excellent guess of the optimal solution. Interior-point methods are not readily warm-started and are not as efficiently warm-started as are active-set methods, despite recent efforts in the area =-=[2, 11, 13]-=-. PDE-constrained optimization problems give rise to huge linear systems that must be solved at every iteration. In particular, the linear systems arising from any optimization method can no longer be... |

85 | Newton’s method for large bound-constrained optimization problems, preprint
- Lin, Moré
- 1999
(Show Context)
Citation Context ... � c(x) − µAT (x)g(x) �T µ y + 2 2 gT (x)g(x) from which the solution d = −µg(x)+A(x)y of RLP(x, µ) may be recovered. The significance is that there are excellent methods [7, 15, 21, 23] and software =-=[8, 19]-=- for solving large-scale BQPs. Having solved RLP(x, µ)—or its dual, BQP(x, µ)—to find an initial step dLP and its associated active set A, our algorithm solves the resulting EQP(x, A) to find a second... |

79 |
On the solution of large quadratic programming problems with bound constraints
- Moré, Toraldo
- 1991
(Show Context)
Citation Context ...to y ≥ 0, 1 2yT AT (x)A(x)y + � c(x) − µAT (x)g(x) �T µ y + 2 2 gT (x)g(x) from which the solution d = −µg(x)+A(x)y of RLP(x, µ) may be recovered. The significance is that there are excellent methods =-=[7, 15, 21, 23]-=- and software [8, 19] for solving large-scale BQPs. Having solved RLP(x, µ)—or its dual, BQP(x, µ)—to find an initial step dLP and its associated active set A, our algorithm solves the resulting EQP(x... |

66 | Global convergence of a class of trust region algorithms for optimization with simple bounds - Toint - 1988 |

57 |
Nonlinear perturbations of linear programs
- Mangasarian, Meyer
- 1979
(Show Context)
Citation Context ...interesting history. In particular, RLP(x, µ) is a Tikhonov regularization [27] of LP(x), for which the solution of the former is the least ℓ2-norm solution to the latter for all µ sufficiently large =-=[22]-=-. See [29] for further references. Vitally, since the Hessian of the objective is diagonal, the dual of RLP(x, µ) is a bound-constrained quadratic program (BQP) [16, 21] in the dual variables y, (BQP(... |

50 | On the global convergence of an SLP-filter algorithm that takes EQP steps
- Chin, Fletcher
- 2003
(Show Context)
Citation Context ...y the optimal active set. Next, we describe our globalization strategy and provide a formal statement of our new algorithm. 2.1 Sequential Linear-Quadratic Programming Methods We believe SLQP methods =-=[3,4,6,24]-=- are well suited for large-scale nonlinear programming, simply because each of the key steps involved—the solution an LP or EQP, together with its attendant linear algebra—has already been tuned to co... |

49 | Knitro: An integrated package for nonlinear optimization
- Byrd, Nocedal, et al.
- 2006
(Show Context)
Citation Context ...e active set for one subproblem is a good prediction for that of the next (and thus allows efficient “warm starts”). Extensive numerical experience with the SLIQUE and KNITRO-ACTIVE software packages =-=[3, 5]-=- has indicated that indeed in some cases warm-start strategies may be inefficient for this reason. One way to avoid this drawback is to use a trust-region norm that does not have multiple faces. Then,... |

46 |
Uniqueness of solution in linear programming
- Mangasarian
- 1979
(Show Context)
Citation Context ...atter for all µ sufficiently large [22]. See [29] for further references. Vitally, since the Hessian of the objective is diagonal, the dual of RLP(x, µ) is a bound-constrained quadratic program (BQP) =-=[16, 21]-=- in the dual variables y, (BQP(x, µ)) � minimize y subject to y ≥ 0, 1 2yT AT (x)A(x)y + � c(x) − µAT (x)g(x) �T µ y + 2 2 gT (x)g(x) from which the solution d = −µg(x)+A(x)y of RLP(x, µ) may be recov... |

42 | An algorithm for nonlinear optimization using linear programming and equality constrained subproblems
- Byrd, Gould, et al.
(Show Context)
Citation Context ...y the optimal active set. Next, we describe our globalization strategy and provide a formal statement of our new algorithm. 2.1 Sequential Linear-Quadratic Programming Methods We believe SLQP methods =-=[3,4,6,24]-=- are well suited for large-scale nonlinear programming, simply because each of the key steps involved—the solution an LP or EQP, together with its attendant linear algebra—has already been tuned to co... |

31 |
de la Maza. Nonlinear programming and nonsmooth optimization by successive linear programming
- Fletcher, Sainz
- 1989
(Show Context)
Citation Context ...y the optimal active set. Next, we describe our globalization strategy and provide a formal statement of our new algorithm. 2.1 Sequential Linear-Quadratic Programming Methods We believe SLQP methods =-=[3,4,6,24]-=- are well suited for large-scale nonlinear programming, simply because each of the key steps involved—the solution an LP or EQP, together with its attendant linear algebra—has already been tuned to co... |

20 |
On the convergence of successive linear-quadratic programming algorithms
- Byrd, Gould, et al.
- 2004
(Show Context)
Citation Context |

14 | A superlinearly convergent sequential quadratically constrained quadratic programming algorithm for degenerate nonlinear programming
- Anitescu
(Show Context)
Citation Context ... appropriate, but the downside is that TR(x, ρ) is no longer an LP and may be more difficult to solve—there has been some work for the ℓ2-norm case using semi-definite and cone programming techniques =-=[1, 12, 17, 18, 20, 26]-=-, but to our knowledge the size of problems that are amenable to such techniques fall far short of those currently possible for LP subproblems. A second difficulty, common to many trust-region methods... |

14 |
SQ2P, sequential quadratic constrained quadratic programming
- Kruk, Wolkowicz
- 1998
(Show Context)
Citation Context ... appropriate, but the downside is that TR(x, ρ) is no longer an LP and may be more difficult to solve—there has been some work for the `2-norm case using semi-definite and cone programming techniques =-=[1, 12, 17, 18, 20, 26]-=-, but to our knowledge the size of problems that are amenable to such techniques fall far short of those currently possible for LP subproblems. A second difficulty, common to many trust-region methods... |

13 |
A sequential quadratically constrained quadratic programming method for differentiable convex minimization
- Fukushima, Luo, et al.
(Show Context)
Citation Context ... appropriate, but the downside is that TR(x, ρ) is no longer an LP and may be more difficult to solve—there has been some work for the ℓ2-norm case using semi-definite and cone programming techniques =-=[1, 12, 17, 18, 20, 26]-=-, but to our knowledge the size of problems that are amenable to such techniques fall far short of those currently possible for LP subproblems. A second difficulty, common to many trust-region methods... |

13 | A new unblocking technique to warmstart interior point methods based on sensitivity analysis
- Gondzio, Grothey
- 2006
(Show Context)
Citation Context ...m an excellent guess of the optimal solution. Interior-point methods are not readily warm-started and are not as efficiently warm-started as are active-set methods, despite recent efforts in the area =-=[2, 11, 13]-=-. PDE-constrained optimization problems give rise to huge linear systems that must be solved at every iteration. In particular, the linear systems arising from any optimization method can no longer be... |

9 | On the sequential quadratically constrained quadratic programming methods
- Solodov
(Show Context)
Citation Context |

6 | Semidefinite programming applied to nonlinear programming
- Kruk
- 1996
(Show Context)
Citation Context |

5 | Numerical methods for large-scale nonlinear optimization - Toint - 2005 |

5 |
Locating the least 2-norm solution of linear programs via a path-following methods
- Zhao, Li
- 2002
(Show Context)
Citation Context ...g history. In particular, RLP(x, µ) is a Tikhonov regularization [27] of LP(x), for which the solution of the former is the least ℓ2-norm solution to the latter for all µ sufficiently large [22]. See =-=[29]-=- for further references. Vitally, since the Hessian of the objective is diagonal, the dual of RLP(x, µ) is a bound-constrained quadratic program (BQP) [16, 21] in the dual variables y, (BQP(x, µ)) � m... |

4 |
A quadratic programming bibliography
- Toint
- 2000
(Show Context)
Citation Context ...to y ≥ 0, 1 2yT AT (x)A(x)y + � c(x) − µAT (x)g(x) �T µ y + 2 2 gT (x)g(x) from which the solution d = −µg(x)+A(x)y of RLP(x, µ) may be recovered. The significance is that there are excellent methods =-=[7, 15, 21, 23]-=- and software [8, 19] for solving large-scale BQPs. Having solved RLP(x, µ)—or its dual, BQP(x, µ)—to find an initial step dLP and its associated active set A, our algorithm solves the resulting EQP(x... |

4 |
On the global convergence of an SLP-filter algorithm. Numerical Analysis Report NA/183
- Toint
- 1998
(Show Context)
Citation Context ...d Figure 4: FASTr: Filter Active-Set Trust-Region Methods12 Michael Friedlander, Nick Gould, Sven Leyffer, and Todd Munson 3 Convergence Analysis The global convergence proof is similar to the one in =-=[25]-=- and shows that the iterates generated by FASTr converge to stationary point. In fact, taking RLP steps is sufficient for global convergence. However, we also show that the convergence of the EQP step... |

2 |
SQ 2 P, sequential quadratic constrained quadratic programming
- Kruk, Wolkowicz
- 1998
(Show Context)
Citation Context |

1 |
An adaptive primal-dual warm-start technique for quadratic multiobjective optimization
- Fliege, Heermann, et al.
- 2006
(Show Context)
Citation Context ...m an excellent guess of the optimal solution. Interior-point methods are not readily warm-started and are not as efficiently warm-started as are active-set methods, despite recent efforts in the area =-=[2, 11, 13]-=-. PDE-constrained optimization problems give rise to huge linear systems that must be solved at every iteration. In particular, the linear systems arising from any optimization method can no longer be... |

1 |
Global convergence of a trust region sequential quadratic programming method
- Yamashita, Dan
(Show Context)
Citation Context ...step dSOC are intended to accelerate convergence and thus to be preferred if suitable. We also note that the computation of superficially similar steps has recently been proposed by Yamashita and Dan =-=[28]-=- but with significant differences, the most important being that µ is not used as a control parameter as it is here and that a different globalization strategy is used. 2.4 Globalization and the Filte... |